33.3 Technical issues

For large molecules, we recommend using the density-fitted Hartree-Fock program (DF-HF) for the proceeding Hartree-Fock calculation. Presently the program is not well parallelized over multiple computer nodes. We therefore suggest running DF-HF calculations separately and using the save option to dump the molecular orbitals to disk for the electron correlation calculations. A parallel local density-fitted Hartree-Fock program is being developed and will be available in the near future.

The PNO program stores large data structure in distributed memory through the globalarrays toolkit. The toolkit has memory allocation mechanisms independent of Molpro. When running the PNO program, one should not allocate all available physical memory to Molpro using the memory command in an input file, so that the globalarrays toolkit could allocate sufficient memory when needed. In some cases it is helpful to pass the -G [ga_mem] option in the molpro command line. This allows the allocation of ga_mem megawords of memory for globalarrays at the beginning of execution. Without doing this, globalarrays may crash when the distributed data structure gets large, most likely due to an upstream bug. The total amount of memory used by molpro and globalarrays on a node, in MB, can be estimated from $n_\mathrm{core}(8M_\mathrm{molpro}+1000) + 8M_\mathrm{GA}/n_\mathrm{node}$, where $n_\mathrm{core}$ is the number of cores per node, $n_\mathrm{node}$ is the number of nodes, $M_\mathrm{molpro}$ is the memory set by the memory command in the input file, and $M_\mathrm{GA}$ is the memory set with the -G option. The total amount of memory used shall not exceed available memory on a node, otherwise runtime errors due to insufficient memory are likely.

molpro@molpro.net 2018-10-21